git.oblomov.eu Git - linux-2.6/blob - drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-hw-handler.h"
  11 #include "dm-bio-list.h"
  12 #include "dm-bio-record.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <asm/atomic.h>
  23
  24 #define MESG_STR(x) x, sizeof(x)
  25
  26 /* Path properties */
  27 struct pgpath {
  28         struct list_head list;
  29
  30         struct priority_group *pg;      /* Owning PG */
  31         unsigned fail_count;            /* Cumulative failure count */
  32
  33         struct path path;
  34 };
  35
  36 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  37
  38 /*
  39  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  40  * Each has a path selector which controls which path gets used.
  41  */
  42 struct priority_group {
  43         struct list_head list;
  44
  45         struct multipath *m;            /* Owning multipath instance */
  46         struct path_selector ps;
  47
  48         unsigned pg_num;                /* Reference number */
  49         unsigned bypassed;              /* Temporarily bypass this PG? */
  50
  51         unsigned nr_pgpaths;            /* Number of paths in PG */
  52         struct list_head pgpaths;
  53 };
  54
  55 /* Multipath context */
  56 struct multipath {
  57         struct list_head list;
  58         struct dm_target *ti;
  59
  60         spinlock_t lock;
  61
  62         struct hw_handler hw_handler;
  63         unsigned nr_priority_groups;
  64         struct list_head priority_groups;
  65         unsigned pg_init_required;      /* pg_init needs calling? */
  66         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  67
  68         unsigned nr_valid_paths;        /* Total number of usable paths */
  69         struct pgpath *current_pgpath;
  70         struct priority_group *current_pg;
  71         struct priority_group *next_pg; /* Switch to this PG if set */
  72         unsigned repeat_count;          /* I/Os left before calling PS again */
  73
  74         unsigned queue_io;              /* Must we queue all I/O? */
  75         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  76         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  77
  78         struct work_struct process_queued_ios;
  79         struct bio_list queued_ios;
  80         unsigned queue_size;
  81
  82         struct work_struct trigger_event;
  83
  84         /*
  85          * We must use a mempool of mpath_io structs so that we
  86          * can resubmit bios on error.
  87          */
  88         mempool_t *mpio_pool;
  89 };
  90
  91 /*
  92  * Context information attached to each bio we process.
  93  */
  94 struct mpath_io {
  95         struct pgpath *pgpath;
  96         struct dm_bio_details details;
  97 };
  98
  99 typedef int (*action_fn) (struct pgpath *pgpath);
 100
 101 #define MIN_IOS 256     /* Mempool size */
 102
 103 static kmem_cache_t *_mpio_cache;
 104
 105 struct workqueue_struct *kmultipathd;
 106 static void process_queued_ios(void *data);
 107 static void trigger_event(void *data);
 108
 109
 110 /*-----------------------------------------------
 111  * Allocation routines
 112  *-----------------------------------------------*/
 113
 114 static struct pgpath *alloc_pgpath(void)
 115 {
 116         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
 117
 118         if (pgpath) {
 119                 memset(pgpath, 0, sizeof(*pgpath));
 120                 pgpath->path.is_active = 1;
 121         }
 122
 123         return pgpath;
 124 }
 125
 126 static inline void free_pgpath(struct pgpath *pgpath)
 127 {
 128         kfree(pgpath);
 129 }
 130
 131 static struct priority_group *alloc_priority_group(void)
 132 {
 133         struct priority_group *pg;
 134
 135         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
 136         if (!pg)
 137                 return NULL;
 138
 139         memset(pg, 0, sizeof(*pg));
 140         INIT_LIST_HEAD(&pg->pgpaths);
 141
 142         return pg;
 143 }
 144
 145 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 146 {
 147         struct pgpath *pgpath, *tmp;
 148
 149         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 150                 list_del(&pgpath->list);
 151                 dm_put_device(ti, pgpath->path.dev);
 152                 free_pgpath(pgpath);
 153         }
 154 }
 155
 156 static void free_priority_group(struct priority_group *pg,
 157                                 struct dm_target *ti)
 158 {
 159         struct path_selector *ps = &pg->ps;
 160
 161         if (ps->type) {
 162                 ps->type->destroy(ps);
 163                 dm_put_path_selector(ps->type);
 164         }
 165
 166         free_pgpaths(&pg->pgpaths, ti);
 167         kfree(pg);
 168 }
 169
 170 static struct multipath *alloc_multipath(void)
 171 {
 172         struct multipath *m;
 173
 174         m = kmalloc(sizeof(*m), GFP_KERNEL);
 175         if (m) {
 176                 memset(m, 0, sizeof(*m));
 177                 INIT_LIST_HEAD(&m->priority_groups);
 178                 spin_lock_init(&m->lock);
 179                 m->queue_io = 1;
 180                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
 181                 INIT_WORK(&m->trigger_event, trigger_event, m);
 182                 m->mpio_pool = mempool_create(MIN_IOS, mempool_alloc_slab,
 183                                               mempool_free_slab, _mpio_cache);
 184                 if (!m->mpio_pool) {
 185                         kfree(m);
 186                         return NULL;
 187                 }
 188         }
 189
 190         return m;
 191 }
 192
 193 static void free_multipath(struct multipath *m)
 194 {
 195         struct priority_group *pg, *tmp;
 196         struct hw_handler *hwh = &m->hw_handler;
 197
 198         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 199                 list_del(&pg->list);
 200                 free_priority_group(pg, m->ti);
 201         }
 202
 203         if (hwh->type) {
 204                 hwh->type->destroy(hwh);
 205                 dm_put_hw_handler(hwh->type);
 206         }
 207
 208         mempool_destroy(m->mpio_pool);
 209         kfree(m);
 210 }
 211
 212
 213 /*-----------------------------------------------
 214  * Path selection
 215  *-----------------------------------------------*/
 216
 217 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 218 {
 219         struct hw_handler *hwh = &m->hw_handler;
 220
 221         m->current_pg = pgpath->pg;
 222
 223         /* Must we initialise the PG first, and queue I/O till it's ready? */
 224         if (hwh->type && hwh->type->pg_init) {
 225                 m->pg_init_required = 1;
 226                 m->queue_io = 1;
 227         } else {
 228                 m->pg_init_required = 0;
 229                 m->queue_io = 0;
 230         }
 231 }
 232
 233 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 234 {
 235         struct path *path;
 236
 237         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 238         if (!path)
 239                 return -ENXIO;
 240
 241         m->current_pgpath = path_to_pgpath(path);
 242
 243         if (m->current_pg != pg)
 244                 __switch_pg(m, m->current_pgpath);
 245
 246         return 0;
 247 }
 248
 249 static void __choose_pgpath(struct multipath *m)
 250 {
 251         struct priority_group *pg;
 252         unsigned bypassed = 1;
 253
 254         if (!m->nr_valid_paths)
 255                 goto failed;
 256
 257         /* Were we instructed to switch PG? */
 258         if (m->next_pg) {
 259                 pg = m->next_pg;
 260                 m->next_pg = NULL;
 261                 if (!__choose_path_in_pg(m, pg))
 262                         return;
 263         }
 264
 265         /* Don't change PG until it has no remaining paths */
 266         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 267                 return;
 268
 269         /*
 270          * Loop through priority groups until we find a valid path.
 271          * First time we skip PGs marked 'bypassed'.
 272          * Second time we only try the ones we skipped.
 273          */
 274         do {
 275                 list_for_each_entry(pg, &m->priority_groups, list) {
 276                         if (pg->bypassed == bypassed)
 277                                 continue;
 278                         if (!__choose_path_in_pg(m, pg))
 279                                 return;
 280                 }
 281         } while (bypassed--);
 282
 283 failed:
 284         m->current_pgpath = NULL;
 285         m->current_pg = NULL;
 286 }
 287
 288 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 289                   unsigned was_queued)
 290 {
 291         int r = 1;
 292         unsigned long flags;
 293         struct pgpath *pgpath;
 294
 295         spin_lock_irqsave(&m->lock, flags);
 296
 297         /* Do we need to select a new pgpath? */
 298         if (!m->current_pgpath ||
 299             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 300                 __choose_pgpath(m);
 301
 302         pgpath = m->current_pgpath;
 303
 304         if (was_queued)
 305                 m->queue_size--;
 306
 307         if ((pgpath && m->queue_io) ||
 308             (!pgpath && m->queue_if_no_path)) {
 309                 /* Queue for the daemon to resubmit */
 310                 bio_list_add(&m->queued_ios, bio);
 311                 m->queue_size++;
 312                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 313                     !m->queue_io)
 314                         queue_work(kmultipathd, &m->process_queued_ios);
 315                 pgpath = NULL;
 316                 r = 0;
 317         } else if (!pgpath)
 318                 r = -EIO;               /* Failed */
 319         else
 320                 bio->bi_bdev = pgpath->path.dev->bdev;
 321
 322         mpio->pgpath = pgpath;
 323
 324         spin_unlock_irqrestore(&m->lock, flags);
 325
 326         return r;
 327 }
 328
 329 /*
 330  * If we run out of usable paths, should we queue I/O or error it?
 331  */
 332 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path)
 333 {
 334         unsigned long flags;
 335
 336         spin_lock_irqsave(&m->lock, flags);
 337
 338         m->saved_queue_if_no_path = m->queue_if_no_path;
 339         m->queue_if_no_path = queue_if_no_path;
 340         if (!m->queue_if_no_path && m->queue_size)
 341                 queue_work(kmultipathd, &m->process_queued_ios);
 342
 343         spin_unlock_irqrestore(&m->lock, flags);
 344
 345         return 0;
 346 }
 347
 348 /*-----------------------------------------------------------------
 349  * The multipath daemon is responsible for resubmitting queued ios.
 350  *---------------------------------------------------------------*/
 351
 352 static void dispatch_queued_ios(struct multipath *m)
 353 {
 354         int r;
 355         unsigned long flags;
 356         struct bio *bio = NULL, *next;
 357         struct mpath_io *mpio;
 358         union map_info *info;
 359
 360         spin_lock_irqsave(&m->lock, flags);
 361         bio = bio_list_get(&m->queued_ios);
 362         spin_unlock_irqrestore(&m->lock, flags);
 363
 364         while (bio) {
 365                 next = bio->bi_next;
 366                 bio->bi_next = NULL;
 367
 368                 info = dm_get_mapinfo(bio);
 369                 mpio = info->ptr;
 370
 371                 r = map_io(m, bio, mpio, 1);
 372                 if (r < 0)
 373                         bio_endio(bio, bio->bi_size, r);
 374                 else if (r == 1)
 375                         generic_make_request(bio);
 376
 377                 bio = next;
 378         }
 379 }
 380
 381 static void process_queued_ios(void *data)
 382 {
 383         struct multipath *m = (struct multipath *) data;
 384         struct hw_handler *hwh = &m->hw_handler;
 385         struct pgpath *pgpath = NULL;
 386         unsigned init_required = 0, must_queue = 1;
 387         unsigned long flags;
 388
 389         spin_lock_irqsave(&m->lock, flags);
 390
 391         if (!m->queue_size)
 392                 goto out;
 393
 394         if (!m->current_pgpath)
 395                 __choose_pgpath(m);
 396
 397         pgpath = m->current_pgpath;
 398
 399         if ((pgpath && !m->queue_io) ||
 400             (!pgpath && !m->queue_if_no_path))
 401                 must_queue = 0;
 402
 403         if (m->pg_init_required && !m->pg_init_in_progress) {
 404                 m->pg_init_required = 0;
 405                 m->pg_init_in_progress = 1;
 406                 init_required = 1;
 407         }
 408
 409 out:
 410         spin_unlock_irqrestore(&m->lock, flags);
 411
 412         if (init_required)
 413                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 414
 415         if (!must_queue)
 416                 dispatch_queued_ios(m);
 417 }
 418
 419 /*
 420  * An event is triggered whenever a path is taken out of use.
 421  * Includes path failure and PG bypass.
 422  */
 423 static void trigger_event(void *data)
 424 {
 425         struct multipath *m = (struct multipath *) data;
 426
 427         dm_table_event(m->ti->table);
 428 }
 429
 430 /*-----------------------------------------------------------------
 431  * Constructor/argument parsing:
 432  * <#multipath feature args> [<arg>]*
 433  * <#hw_handler args> [hw_handler [<arg>]*]
 434  * <#priority groups>
 435  * <initial priority group>
 436  *     [<selector> <#selector args> [<arg>]*
 437  *      <#paths> <#per-path selector args>
 438  *         [<path> [<arg>]* ]+ ]+
 439  *---------------------------------------------------------------*/
 440 struct param {
 441         unsigned min;
 442         unsigned max;
 443         char *error;
 444 };
 445
 446 #define ESTR(s) ("dm-multipath: " s)
 447
 448 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 449 {
 450         if (!str ||
 451             (sscanf(str, "%u", v) != 1) ||
 452             (*v < param->min) ||
 453             (*v > param->max)) {
 454                 *error = param->error;
 455                 return -EINVAL;
 456         }
 457
 458         return 0;
 459 }
 460
 461 struct arg_set {
 462         unsigned argc;
 463         char **argv;
 464 };
 465
 466 static char *shift(struct arg_set *as)
 467 {
 468         char *r;
 469
 470         if (as->argc) {
 471                 as->argc--;
 472                 r = *as->argv;
 473                 as->argv++;
 474                 return r;
 475         }
 476
 477         return NULL;
 478 }
 479
 480 static void consume(struct arg_set *as, unsigned n)
 481 {
 482         BUG_ON (as->argc < n);
 483         as->argc -= n;
 484         as->argv += n;
 485 }
 486
 487 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 488                                struct dm_target *ti)
 489 {
 490         int r;
 491         struct path_selector_type *pst;
 492         unsigned ps_argc;
 493
 494         static struct param _params[] = {
 495                 {0, 1024, ESTR("invalid number of path selector args")},
 496         };
 497
 498         pst = dm_get_path_selector(shift(as));
 499         if (!pst) {
 500                 ti->error = ESTR("unknown path selector type");
 501                 return -EINVAL;
 502         }
 503
 504         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 505         if (r)
 506                 return -EINVAL;
 507
 508         r = pst->create(&pg->ps, ps_argc, as->argv);
 509         if (r) {
 510                 dm_put_path_selector(pst);
 511                 ti->error = ESTR("path selector constructor failed");
 512                 return r;
 513         }
 514
 515         pg->ps.type = pst;
 516         consume(as, ps_argc);
 517
 518         return 0;
 519 }
 520
 521 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 522                                struct dm_target *ti)
 523 {
 524         int r;
 525         struct pgpath *p;
 526
 527         /* we need at least a path arg */
 528         if (as->argc < 1) {
 529                 ti->error = ESTR("no device given");
 530                 return NULL;
 531         }
 532
 533         p = alloc_pgpath();
 534         if (!p)
 535                 return NULL;
 536
 537         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 538                           dm_table_get_mode(ti->table), &p->path.dev);
 539         if (r) {
 540                 ti->error = ESTR("error getting device");
 541                 goto bad;
 542         }
 543
 544         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 545         if (r) {
 546                 dm_put_device(ti, p->path.dev);
 547                 goto bad;
 548         }
 549
 550         return p;
 551
 552  bad:
 553         free_pgpath(p);
 554         return NULL;
 555 }
 556
 557 static struct priority_group *parse_priority_group(struct arg_set *as,
 558                                                    struct multipath *m,
 559                                                    struct dm_target *ti)
 560 {
 561         static struct param _params[] = {
 562                 {1, 1024, ESTR("invalid number of paths")},
 563                 {0, 1024, ESTR("invalid number of selector args")}
 564         };
 565
 566         int r;
 567         unsigned i, nr_selector_args, nr_params;
 568         struct priority_group *pg;
 569
 570         if (as->argc < 2) {
 571                 as->argc = 0;
 572                 ti->error = ESTR("not enough priority group aruments");
 573                 return NULL;
 574         }
 575
 576         pg = alloc_priority_group();
 577         if (!pg) {
 578                 ti->error = ESTR("couldn't allocate priority group");
 579                 return NULL;
 580         }
 581         pg->m = m;
 582
 583         r = parse_path_selector(as, pg, ti);
 584         if (r)
 585                 goto bad;
 586
 587         /*
 588          * read the paths
 589          */
 590         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 591         if (r)
 592                 goto bad;
 593
 594         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 595         if (r)
 596                 goto bad;
 597
 598         nr_params = 1 + nr_selector_args;
 599         for (i = 0; i < pg->nr_pgpaths; i++) {
 600                 struct pgpath *pgpath;
 601                 struct arg_set path_args;
 602
 603                 if (as->argc < nr_params)
 604                         goto bad;
 605
 606                 path_args.argc = nr_params;
 607                 path_args.argv = as->argv;
 608
 609                 pgpath = parse_path(&path_args, &pg->ps, ti);
 610                 if (!pgpath)
 611                         goto bad;
 612
 613                 pgpath->pg = pg;
 614                 list_add_tail(&pgpath->list, &pg->pgpaths);
 615                 consume(as, nr_params);
 616         }
 617
 618         return pg;
 619
 620  bad:
 621         free_priority_group(pg, ti);
 622         return NULL;
 623 }
 624
 625 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
 626                             struct dm_target *ti)
 627 {
 628         int r;
 629         struct hw_handler_type *hwht;
 630         unsigned hw_argc;
 631
 632         static struct param _params[] = {
 633                 {0, 1024, ESTR("invalid number of hardware handler args")},
 634         };
 635
 636         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 637         if (r)
 638                 return -EINVAL;
 639
 640         if (!hw_argc)
 641                 return 0;
 642
 643         hwht = dm_get_hw_handler(shift(as));
 644         if (!hwht) {
 645                 ti->error = ESTR("unknown hardware handler type");
 646                 return -EINVAL;
 647         }
 648
 649         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 650         if (r) {
 651                 dm_put_hw_handler(hwht);
 652                 ti->error = ESTR("hardware handler constructor failed");
 653                 return r;
 654         }
 655
 656         m->hw_handler.type = hwht;
 657         consume(as, hw_argc - 1);
 658
 659         return 0;
 660 }
 661
 662 static int parse_features(struct arg_set *as, struct multipath *m,
 663                           struct dm_target *ti)
 664 {
 665         int r;
 666         unsigned argc;
 667
 668         static struct param _params[] = {
 669                 {0, 1, ESTR("invalid number of feature args")},
 670         };
 671
 672         r = read_param(_params, shift(as), &argc, &ti->error);
 673         if (r)
 674                 return -EINVAL;
 675
 676         if (!argc)
 677                 return 0;
 678
 679         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 680                 return queue_if_no_path(m, 1);
 681         else {
 682                 ti->error = "Unrecognised multipath feature request";
 683                 return -EINVAL;
 684         }
 685 }
 686
 687 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 688                          char **argv)
 689 {
 690         /* target parameters */
 691         static struct param _params[] = {
 692                 {1, 1024, ESTR("invalid number of priority groups")},
 693                 {1, 1024, ESTR("invalid initial priority group number")},
 694         };
 695
 696         int r;
 697         struct multipath *m;
 698         struct arg_set as;
 699         unsigned pg_count = 0;
 700         unsigned next_pg_num;
 701
 702         as.argc = argc;
 703         as.argv = argv;
 704
 705         m = alloc_multipath();
 706         if (!m) {
 707                 ti->error = ESTR("can't allocate multipath");
 708                 return -EINVAL;
 709         }
 710
 711         r = parse_features(&as, m, ti);
 712         if (r)
 713                 goto bad;
 714
 715         r = parse_hw_handler(&as, m, ti);
 716         if (r)
 717                 goto bad;
 718
 719         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 720         if (r)
 721                 goto bad;
 722
 723         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 724         if (r)
 725                 goto bad;
 726
 727         /* parse the priority groups */
 728         while (as.argc) {
 729                 struct priority_group *pg;
 730
 731                 pg = parse_priority_group(&as, m, ti);
 732                 if (!pg) {
 733                         r = -EINVAL;
 734                         goto bad;
 735                 }
 736
 737                 m->nr_valid_paths += pg->nr_pgpaths;
 738                 list_add_tail(&pg->list, &m->priority_groups);
 739                 pg_count++;
 740                 pg->pg_num = pg_count;
 741                 if (!--next_pg_num)
 742                         m->next_pg = pg;
 743         }
 744
 745         if (pg_count != m->nr_priority_groups) {
 746                 ti->error = ESTR("priority group count mismatch");
 747                 r = -EINVAL;
 748                 goto bad;
 749         }
 750
 751         ti->private = m;
 752         m->ti = ti;
 753
 754         return 0;
 755
 756  bad:
 757         free_multipath(m);
 758         return r;
 759 }
 760
 761 static void multipath_dtr(struct dm_target *ti)
 762 {
 763         struct multipath *m = (struct multipath *) ti->private;
 764
 765         flush_workqueue(kmultipathd);
 766         free_multipath(m);
 767 }
 768
 769 /*
 770  * Map bios, recording original fields for later in case we have to resubmit
 771  */
 772 static int multipath_map(struct dm_target *ti, struct bio *bio,
 773                          union map_info *map_context)
 774 {
 775         int r;
 776         struct mpath_io *mpio;
 777         struct multipath *m = (struct multipath *) ti->private;
 778
 779         if (bio_barrier(bio))
 780                 return -EOPNOTSUPP;
 781
 782         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 783         dm_bio_record(&mpio->details, bio);
 784
 785         map_context->ptr = mpio;
 786         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 787         r = map_io(m, bio, mpio, 0);
 788         if (r < 0)
 789                 mempool_free(mpio, m->mpio_pool);
 790
 791         return r;
 792 }
 793
 794 /*
 795  * Take a path out of use.
 796  */
 797 static int fail_path(struct pgpath *pgpath)
 798 {
 799         unsigned long flags;
 800         struct multipath *m = pgpath->pg->m;
 801
 802         spin_lock_irqsave(&m->lock, flags);
 803
 804         if (!pgpath->path.is_active)
 805                 goto out;
 806
 807         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
 808
 809         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 810         pgpath->path.is_active = 0;
 811         pgpath->fail_count++;
 812
 813         m->nr_valid_paths--;
 814
 815         if (pgpath == m->current_pgpath)
 816                 m->current_pgpath = NULL;
 817
 818         queue_work(kmultipathd, &m->trigger_event);
 819
 820 out:
 821         spin_unlock_irqrestore(&m->lock, flags);
 822
 823         return 0;
 824 }
 825
 826 /*
 827  * Reinstate a previously-failed path
 828  */
 829 static int reinstate_path(struct pgpath *pgpath)
 830 {
 831         int r = 0;
 832         unsigned long flags;
 833         struct multipath *m = pgpath->pg->m;
 834
 835         spin_lock_irqsave(&m->lock, flags);
 836
 837         if (pgpath->path.is_active)
 838                 goto out;
 839
 840         if (!pgpath->pg->ps.type) {
 841                 DMWARN("Reinstate path not supported by path selector %s",
 842                        pgpath->pg->ps.type->name);
 843                 r = -EINVAL;
 844                 goto out;
 845         }
 846
 847         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 848         if (r)
 849                 goto out;
 850
 851         pgpath->path.is_active = 1;
 852
 853         m->current_pgpath = NULL;
 854         if (!m->nr_valid_paths++ && m->queue_size)
 855                 queue_work(kmultipathd, &m->process_queued_ios);
 856
 857         queue_work(kmultipathd, &m->trigger_event);
 858
 859 out:
 860         spin_unlock_irqrestore(&m->lock, flags);
 861
 862         return r;
 863 }
 864
 865 /*
 866  * Fail or reinstate all paths that match the provided struct dm_dev.
 867  */
 868 static int action_dev(struct multipath *m, struct dm_dev *dev,
 869                       action_fn action)
 870 {
 871         int r = 0;
 872         struct pgpath *pgpath;
 873         struct priority_group *pg;
 874
 875         list_for_each_entry(pg, &m->priority_groups, list) {
 876                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 877                         if (pgpath->path.dev == dev)
 878                                 r = action(pgpath);
 879                 }
 880         }
 881
 882         return r;
 883 }
 884
 885 /*
 886  * Temporarily try to avoid having to use the specified PG
 887  */
 888 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 889                       int bypassed)
 890 {
 891         unsigned long flags;
 892
 893         spin_lock_irqsave(&m->lock, flags);
 894
 895         pg->bypassed = bypassed;
 896         m->current_pgpath = NULL;
 897         m->current_pg = NULL;
 898
 899         spin_unlock_irqrestore(&m->lock, flags);
 900
 901         queue_work(kmultipathd, &m->trigger_event);
 902 }
 903
 904 /*
 905  * Switch to using the specified PG from the next I/O that gets mapped
 906  */
 907 static int switch_pg_num(struct multipath *m, const char *pgstr)
 908 {
 909         struct priority_group *pg;
 910         unsigned pgnum;
 911         unsigned long flags;
 912
 913         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 914             (pgnum > m->nr_priority_groups)) {
 915                 DMWARN("invalid PG number supplied to switch_pg_num");
 916                 return -EINVAL;
 917         }
 918
 919         spin_lock_irqsave(&m->lock, flags);
 920         list_for_each_entry(pg, &m->priority_groups, list) {
 921                 pg->bypassed = 0;
 922                 if (--pgnum)
 923                         continue;
 924
 925                 m->current_pgpath = NULL;
 926                 m->current_pg = NULL;
 927                 m->next_pg = pg;
 928         }
 929         spin_unlock_irqrestore(&m->lock, flags);
 930
 931         queue_work(kmultipathd, &m->trigger_event);
 932         return 0;
 933 }
 934
 935 /*
 936  * Set/clear bypassed status of a PG.
 937  * PGs are numbered upwards from 1 in the order they were declared.
 938  */
 939 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 940 {
 941         struct priority_group *pg;
 942         unsigned pgnum;
 943
 944         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 945             (pgnum > m->nr_priority_groups)) {
 946                 DMWARN("invalid PG number supplied to bypass_pg");
 947                 return -EINVAL;
 948         }
 949
 950         list_for_each_entry(pg, &m->priority_groups, list) {
 951                 if (!--pgnum)
 952                         break;
 953         }
 954
 955         bypass_pg(m, pg, bypassed);
 956         return 0;
 957 }
 958
 959 /*
 960  * pg_init must call this when it has completed its initialisation
 961  */
 962 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 963 {
 964         struct pgpath *pgpath = path_to_pgpath(path);
 965         struct priority_group *pg = pgpath->pg;
 966         struct multipath *m = pg->m;
 967         unsigned long flags;
 968
 969         /* We insist on failing the path if the PG is already bypassed. */
 970         if (err_flags && pg->bypassed)
 971                 err_flags |= MP_FAIL_PATH;
 972
 973         if (err_flags & MP_FAIL_PATH)
 974                 fail_path(pgpath);
 975
 976         if (err_flags & MP_BYPASS_PG)
 977                 bypass_pg(m, pg, 1);
 978
 979         spin_lock_irqsave(&m->lock, flags);
 980         if (err_flags) {
 981                 m->current_pgpath = NULL;
 982                 m->current_pg = NULL;
 983         } else if (!m->pg_init_required)
 984                 m->queue_io = 0;
 985
 986         m->pg_init_in_progress = 0;
 987         queue_work(kmultipathd, &m->process_queued_ios);
 988         spin_unlock_irqrestore(&m->lock, flags);
 989 }
 990
 991 /*
 992  * end_io handling
 993  */
 994 static int do_end_io(struct multipath *m, struct bio *bio,
 995                      int error, struct mpath_io *mpio)
 996 {
 997         struct hw_handler *hwh = &m->hw_handler;
 998         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
 999
1000         if (!error)
1001                 return 0;       /* I/O complete */
1002
1003         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1004                 return error;
1005
1006         if (error == -EOPNOTSUPP)
1007                 return error;
1008
1009         spin_lock(&m->lock);
1010         if (!m->nr_valid_paths) {
1011                 if (!m->queue_if_no_path) {
1012                         spin_unlock(&m->lock);
1013                         return -EIO;
1014                 } else {
1015                         spin_unlock(&m->lock);
1016                         goto requeue;
1017                 }
1018         }
1019         spin_unlock(&m->lock);
1020
1021         if (hwh->type && hwh->type->error)
1022                 err_flags = hwh->type->error(hwh, bio);
1023
1024         if (mpio->pgpath) {
1025                 if (err_flags & MP_FAIL_PATH)
1026                         fail_path(mpio->pgpath);
1027
1028                 if (err_flags & MP_BYPASS_PG)
1029                         bypass_pg(m, mpio->pgpath->pg, 1);
1030         }
1031
1032         if (err_flags & MP_ERROR_IO)
1033                 return -EIO;
1034
1035       requeue:
1036         dm_bio_restore(&mpio->details, bio);
1037
1038         /* queue for the daemon to resubmit or fail */
1039         spin_lock(&m->lock);
1040         bio_list_add(&m->queued_ios, bio);
1041         m->queue_size++;
1042         if (!m->queue_io)
1043                 queue_work(kmultipathd, &m->process_queued_ios);
1044         spin_unlock(&m->lock);
1045
1046         return 1;       /* io not complete */
1047 }
1048
1049 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1050                             int error, union map_info *map_context)
1051 {
1052         struct multipath *m = (struct multipath *) ti->private;
1053         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1054         struct pgpath *pgpath = mpio->pgpath;
1055         struct path_selector *ps;
1056         int r;
1057
1058         r  = do_end_io(m, bio, error, mpio);
1059         if (pgpath) {
1060                 ps = &pgpath->pg->ps;
1061                 if (ps->type->end_io)
1062                         ps->type->end_io(ps, &pgpath->path);
1063         }
1064         if (r <= 0)
1065                 mempool_free(mpio, m->mpio_pool);
1066
1067         return r;
1068 }
1069
1070 /*
1071  * Suspend can't complete until all the I/O is processed so if
1072  * the last path fails we must error any remaining I/O.
1073  * Note that if the freeze_bdev fails while suspending, the
1074  * queue_if_no_path state is lost - userspace should reset it.
1075  */
1076 static void multipath_presuspend(struct dm_target *ti)
1077 {
1078         struct multipath *m = (struct multipath *) ti->private;
1079
1080         queue_if_no_path(m, 0);
1081 }
1082
1083 /*
1084  * Restore the queue_if_no_path setting.
1085  */
1086 static void multipath_resume(struct dm_target *ti)
1087 {
1088         struct multipath *m = (struct multipath *) ti->private;
1089         unsigned long flags;
1090
1091         spin_lock_irqsave(&m->lock, flags);
1092         m->queue_if_no_path = m->saved_queue_if_no_path;
1093         spin_unlock_irqrestore(&m->lock, flags);
1094 }
1095
1096 /*
1097  * Info output has the following format:
1098  * num_multipath_feature_args [multipath_feature_args]*
1099  * num_handler_status_args [handler_status_args]*
1100  * num_groups init_group_number
1101  *            [A|D|E num_ps_status_args [ps_status_args]*
1102  *             num_paths num_selector_args
1103  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1104  *
1105  * Table output has the following format (identical to the constructor string):
1106  * num_feature_args [features_args]*
1107  * num_handler_args hw_handler [hw_handler_args]*
1108  * num_groups init_group_number
1109  *     [priority selector-name num_ps_args [ps_args]*
1110  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1111  */
1112 static int multipath_status(struct dm_target *ti, status_type_t type,
1113                             char *result, unsigned int maxlen)
1114 {
1115         int sz = 0;
1116         unsigned long flags;
1117         struct multipath *m = (struct multipath *) ti->private;
1118         struct hw_handler *hwh = &m->hw_handler;
1119         struct priority_group *pg;
1120         struct pgpath *p;
1121         unsigned pg_num;
1122         char state;
1123
1124         spin_lock_irqsave(&m->lock, flags);
1125
1126         /* Features */
1127         if (type == STATUSTYPE_INFO)
1128                 DMEMIT("1 %u ", m->queue_size);
1129         else if (m->queue_if_no_path)
1130                 DMEMIT("1 queue_if_no_path ");
1131         else
1132                 DMEMIT("0 ");
1133
1134         if (hwh->type && hwh->type->status)
1135                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1136         else if (!hwh->type || type == STATUSTYPE_INFO)
1137                 DMEMIT("0 ");
1138         else
1139                 DMEMIT("1 %s ", hwh->type->name);
1140
1141         DMEMIT("%u ", m->nr_priority_groups);
1142
1143         if (m->next_pg)
1144                 pg_num = m->next_pg->pg_num;
1145         else if (m->current_pg)
1146                 pg_num = m->current_pg->pg_num;
1147         else
1148                         pg_num = 1;
1149
1150         DMEMIT("%u ", pg_num);
1151
1152         switch (type) {
1153         case STATUSTYPE_INFO:
1154                 list_for_each_entry(pg, &m->priority_groups, list) {
1155                         if (pg->bypassed)
1156                                 state = 'D';    /* Disabled */
1157                         else if (pg == m->current_pg)
1158                                 state = 'A';    /* Currently Active */
1159                         else
1160                                 state = 'E';    /* Enabled */
1161
1162                         DMEMIT("%c ", state);
1163
1164                         if (pg->ps.type->status)
1165                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1166                                                           result + sz,
1167                                                           maxlen - sz);
1168                         else
1169                                 DMEMIT("0 ");
1170
1171                         DMEMIT("%u %u ", pg->nr_pgpaths,
1172                                pg->ps.type->info_args);
1173
1174                         list_for_each_entry(p, &pg->pgpaths, list) {
1175                                 DMEMIT("%s %s %u ", p->path.dev->name,
1176                                        p->path.is_active ? "A" : "F",
1177                                        p->fail_count);
1178                                 if (pg->ps.type->status)
1179                                         sz += pg->ps.type->status(&pg->ps,
1180                                               &p->path, type, result + sz,
1181                                               maxlen - sz);
1182                         }
1183                 }
1184                 break;
1185
1186         case STATUSTYPE_TABLE:
1187                 list_for_each_entry(pg, &m->priority_groups, list) {
1188                         DMEMIT("%s ", pg->ps.type->name);
1189
1190                         if (pg->ps.type->status)
1191                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1192                                                           result + sz,
1193                                                           maxlen - sz);
1194                         else
1195                                 DMEMIT("0 ");
1196
1197                         DMEMIT("%u %u ", pg->nr_pgpaths,
1198                                pg->ps.type->table_args);
1199
1200                         list_for_each_entry(p, &pg->pgpaths, list) {
1201                                 DMEMIT("%s ", p->path.dev->name);
1202                                 if (pg->ps.type->status)
1203                                         sz += pg->ps.type->status(&pg->ps,
1204                                               &p->path, type, result + sz,
1205                                               maxlen - sz);
1206                         }
1207                 }
1208                 break;
1209         }
1210
1211         spin_unlock_irqrestore(&m->lock, flags);
1212
1213         return 0;
1214 }
1215
1216 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1217 {
1218         int r;
1219         struct dm_dev *dev;
1220         struct multipath *m = (struct multipath *) ti->private;
1221         action_fn action;
1222
1223         if (argc == 1) {
1224                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1225                         return queue_if_no_path(m, 1);
1226                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1227                         return queue_if_no_path(m, 0);
1228         }
1229
1230         if (argc != 2)
1231                 goto error;
1232
1233         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1234                 return bypass_pg_num(m, argv[1], 1);
1235         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1236                 return bypass_pg_num(m, argv[1], 0);
1237         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1238                 return switch_pg_num(m, argv[1]);
1239         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1240                 action = reinstate_path;
1241         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1242                 action = fail_path;
1243         else
1244                 goto error;
1245
1246         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1247                           dm_table_get_mode(ti->table), &dev);
1248         if (r) {
1249                 DMWARN("dm-multipath message: error getting device %s",
1250                        argv[1]);
1251                 return -EINVAL;
1252         }
1253
1254         r = action_dev(m, dev, action);
1255
1256         dm_put_device(ti, dev);
1257
1258         return r;
1259
1260 error:
1261         DMWARN("Unrecognised multipath message received.");
1262         return -EINVAL;
1263 }
1264
1265 /*-----------------------------------------------------------------
1266  * Module setup
1267  *---------------------------------------------------------------*/
1268 static struct target_type multipath_target = {
1269         .name = "multipath",
1270         .version = {1, 0, 4},
1271         .module = THIS_MODULE,
1272         .ctr = multipath_ctr,
1273         .dtr = multipath_dtr,
1274         .map = multipath_map,
1275         .end_io = multipath_end_io,
1276         .presuspend = multipath_presuspend,
1277         .resume = multipath_resume,
1278         .status = multipath_status,
1279         .message = multipath_message,
1280 };
1281
1282 static int __init dm_multipath_init(void)
1283 {
1284         int r;
1285
1286         /* allocate a slab for the dm_ios */
1287         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1288                                         0, 0, NULL, NULL);
1289         if (!_mpio_cache)
1290                 return -ENOMEM;
1291
1292         r = dm_register_target(&multipath_target);
1293         if (r < 0) {
1294                 DMERR("%s: register failed %d", multipath_target.name, r);
1295                 kmem_cache_destroy(_mpio_cache);
1296                 return -EINVAL;
1297         }
1298
1299         kmultipathd = create_workqueue("kmpathd");
1300         if (!kmultipathd) {
1301                 DMERR("%s: failed to create workqueue kmpathd",
1302                                 multipath_target.name);
1303                 dm_unregister_target(&multipath_target);
1304                 kmem_cache_destroy(_mpio_cache);
1305                 return -ENOMEM;
1306         }
1307
1308         DMINFO("dm-multipath version %u.%u.%u loaded",
1309                multipath_target.version[0], multipath_target.version[1],
1310                multipath_target.version[2]);
1311
1312         return r;
1313 }
1314
1315 static void __exit dm_multipath_exit(void)
1316 {
1317         int r;
1318
1319         destroy_workqueue(kmultipathd);
1320
1321         r = dm_unregister_target(&multipath_target);
1322         if (r < 0)
1323                 DMERR("%s: target unregister failed %d",
1324                       multipath_target.name, r);
1325         kmem_cache_destroy(_mpio_cache);
1326 }
1327
1328 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1329
1330 module_init(dm_multipath_init);
1331 module_exit(dm_multipath_exit);
1332
1333 MODULE_DESCRIPTION(DM_NAME " multipath target");
1334 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1335 MODULE_LICENSE("GPL");